AI AGENT SKILLS

reset-context-contamination

一个面向 crypto 场景的 Agent 技能。原始说明:Discard the accumulated drafts and framings from this thread and re-derive the task from a clean problem statement.

SKILL.md

SKILL.md


name: reset-context-contamination
description: Discard the accumulated drafts and framings from this thread and re-derive the task from a clean problem statement.
metadata:
version: "0.1.0"


This thread's history has become the problem - prior drafts and framings are anchoring every new attempt. Reset.

First, extract what's worth keeping, in plain factual terms: who is involved, the outcome I want, the hard constraints, the known facts, and - critically - what's already been tried and ruled out, so a fresh derivation doesn't walk back into a dead end.

Then pick the path by severity:

  • Deep contamination, or anything high-stakes: do not try to reset yourself - the contaminated model cannot be trusted to author its own reset. Hand the extracted brief to a fresh subagent, or recommend I start a new session. A genuinely clean context window is the only reliable reset.
  • Mild drift only: re-derive in-thread from the extracted brief, as if seeing the task for the first time. Void the earlier drafts entirely - no referencing, quoting, or patching them. If you are unsure the brief itself is right, show it to me before drafting; otherwise just proceed.